LoraConfig arguments
config = LoraConfig(r=32,
lora_alpha=64,
#target_modules=".decoder.(self_attn|encoder_attn).*(q_proj|v_proj)$",#["q_proj", "v_proj"],
target_modules=["q_proj", "v_proj"],
lora_dropout=0.05,
bias="none")
Training arguments
training_args = TrainingArguments(
output_dir="temp", # change to a repo name of your choice
per_device_train_batch_size=8,
gradient_accumulation_steps=2, # increase by 2x for every 2x decrease in batch size
learning_rate=1e-3,
warmup_steps=10,
max_steps=400, #1500
#evaluation_strategy="steps",
fp16=True,
per_device_eval_batch_size=8,
#generation_max_length=128,
eval_steps=100,
logging_steps=25,
remove_unused_columns=False, # required as the PeftModel forward doesn't have the signature of the wrapped model's forward
label_names=["label"], # same reason as above
)
Training procedure
The following bitsandbytes quantization config was used during training:
- quant_method: bitsandbytes
- load_in_8bit: True
- load_in_4bit: False
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: fp4
- bnb_4bit_use_double_quant: False
- bnb_4bit_compute_dtype: float32
Framework versions
